Overview
- The government-backed model Varya was announced on Friday, June 12, by MeitY Secretary S. Krishnan as part of the India AI Mission and is available as a live demo on Avataar’s site.
- Avataar says it built Varya by distilling Alibaba’s Wan 2.2 and redesigning inference so the model runs in four steps instead of about 50, a change the company reports makes generation much faster.
- Company-reported benchmarks claim a 5-second 720p clip can be produced in about 45 seconds on an NVIDIA H200 and that Varya can be hosted at roughly ₹0.48 (≈$0.005) per second of video.
- Avataar plans to release Varya’s model weights and training data on the government AI Kosh portal so developers can self-host and modify the model, but the startup has not yet published a technical paper or third-party evaluations.
- The launch builds on the India AI Mission’s subsidised compute program for selected startups and is pitched at education, MSMEs, creators and government services while independent benchmarks will determine if the claimed cost and quality gains hold in real use.